Machine Learning for Identifying Atomic Species from Optical Emission Spectra Generated by an Atmospheric Pressure Non-thermal Plasma

Abstract

An automatic recognition method of nine atomic species through ensemble classifiers based on decision trees from experimental data collected by optical emission spectroscopy (OES) is presented. Experimental spectra were obtained from OES of an atmospheric pressure non-thermal plasma (APNTP) generated in parallel circular plates dielectric barrier discharge reactor (DBDR). APNTP’s emission was detected and acquired by a monochromator coupled to a photomultiplier and a data acquisition system. Data were organized in columns as relative intensity versus wavelength to generate a synthetic spectra dataset. The performance categorization of candidate classifiers was assessed using the F1 metric; after that, the grid-search hyperparameter optimization technique allowed the selection of the best combination to construct the final ensemble classifier. After the generation of the synthetic spectra dataset, they were evaluated using parametric statistics with analysis of variance (ANOVA) and non-parametric statistics with Friedman’s tests. Subsequently, the critical distance was obtained by Nemenyi parametric profile, showing the best-classified groups with prediction accuracy of the species between 93 and 100% and a confidence value of 95% in the wavelength range from 200 to 890 nm. Finally, the automatic atomic species recognition test was carried out utilizing a set of nine files, each one corresponding to an experimental spectrum obtained from an APNTP generated in three different argon-oxygen gas mixtures, where Ar I, O I, and O II species with predictions range from 73 to 100% (86.5% mean). Further, the proposed method could be trained to analyze various species generated by some other type of electric discharge.